Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/rdft/scalar/r2cf/hf_16.c @ 56:af97cad61ff0

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add updated build of PortAudio for OSX
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 03 Jan 2017 15:10:52 +0000
parents 2cd0e3b3e1fd
children
rev   line source
Chris@42 1 /*
Chris@42 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42 4 *
Chris@42 5 * This program is free software; you can redistribute it and/or modify
Chris@42 6 * it under the terms of the GNU General Public License as published by
Chris@42 7 * the Free Software Foundation; either version 2 of the License, or
Chris@42 8 * (at your option) any later version.
Chris@42 9 *
Chris@42 10 * This program is distributed in the hope that it will be useful,
Chris@42 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@42 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@42 13 * GNU General Public License for more details.
Chris@42 14 *
Chris@42 15 * You should have received a copy of the GNU General Public License
Chris@42 16 * along with this program; if not, write to the Free Software
Chris@42 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@42 18 *
Chris@42 19 */
Chris@42 20
Chris@42 21 /* This file was automatically generated --- DO NOT EDIT */
Chris@42 22 /* Generated on Sat Jul 30 16:46:26 EDT 2016 */
Chris@42 23
Chris@42 24 #include "codelet-rdft.h"
Chris@42 25
Chris@42 26 #ifdef HAVE_FMA
Chris@42 27
Chris@42 28 /* Generated by: ../../../genfft/gen_hc2hc.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 16 -dit -name hf_16 -include hf.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 174 FP additions, 100 FP multiplications,
Chris@42 32 * (or, 104 additions, 30 multiplications, 70 fused multiply/add),
Chris@42 33 * 95 stack variables, 3 constants, and 64 memory accesses
Chris@42 34 */
Chris@42 35 #include "hf.h"
Chris@42 36
Chris@42 37 static void hf_16(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@42 38 {
Chris@42 39 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@42 40 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@42 41 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@42 42 {
Chris@42 43 INT m;
Chris@42 44 for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 30, MAKE_VOLATILE_STRIDE(32, rs)) {
Chris@42 45 E T2T, T2Q;
Chris@42 46 {
Chris@42 47 E T3A, T3o, T8, T1I, T2w, T35, T2k, T1s, T2p, T36, T2r, T1F, T3k, T1N, T3z;
Chris@42 48 E Tl, T1U, T2W, T1P, Tz, T2g, T30, T25, T11, TB, TE, T2a, T31, T2h, T1e;
Chris@42 49 E TC, T1X, TH, TK, TG, TD, TJ;
Chris@42 50 {
Chris@42 51 E Ta, Td, Tb, T1J, Tg, Tj, Tf, Tc, Ti;
Chris@42 52 {
Chris@42 53 E T1h, T1k, T1n, T2s, T1i, T1q, T1m, T1j, T1p;
Chris@42 54 {
Chris@42 55 E T1, T3n, T3, T6, T2, T5;
Chris@42 56 T1 = cr[0];
Chris@42 57 T3n = ci[0];
Chris@42 58 T3 = cr[WS(rs, 8)];
Chris@42 59 T6 = ci[WS(rs, 8)];
Chris@42 60 T2 = W[14];
Chris@42 61 T5 = W[15];
Chris@42 62 {
Chris@42 63 E T3l, T4, T1g, T3m, T7;
Chris@42 64 T1h = cr[WS(rs, 15)];
Chris@42 65 T1k = ci[WS(rs, 15)];
Chris@42 66 T3l = T2 * T6;
Chris@42 67 T4 = T2 * T3;
Chris@42 68 T1g = W[28];
Chris@42 69 T1n = cr[WS(rs, 7)];
Chris@42 70 T3m = FNMS(T5, T3, T3l);
Chris@42 71 T7 = FMA(T5, T6, T4);
Chris@42 72 T2s = T1g * T1k;
Chris@42 73 T1i = T1g * T1h;
Chris@42 74 T3A = T3n - T3m;
Chris@42 75 T3o = T3m + T3n;
Chris@42 76 T8 = T1 + T7;
Chris@42 77 T1I = T1 - T7;
Chris@42 78 T1q = ci[WS(rs, 7)];
Chris@42 79 T1m = W[12];
Chris@42 80 }
Chris@42 81 T1j = W[29];
Chris@42 82 T1p = W[13];
Chris@42 83 }
Chris@42 84 {
Chris@42 85 E T1u, T1x, T1v, T2l, T1A, T1D, T1z, T1w, T1C;
Chris@42 86 {
Chris@42 87 E T2t, T1l, T2v, T1r, T2u, T1o, T1t;
Chris@42 88 T1u = cr[WS(rs, 3)];
Chris@42 89 T2u = T1m * T1q;
Chris@42 90 T1o = T1m * T1n;
Chris@42 91 T2t = FNMS(T1j, T1h, T2s);
Chris@42 92 T1l = FMA(T1j, T1k, T1i);
Chris@42 93 T2v = FNMS(T1p, T1n, T2u);
Chris@42 94 T1r = FMA(T1p, T1q, T1o);
Chris@42 95 T1x = ci[WS(rs, 3)];
Chris@42 96 T1t = W[4];
Chris@42 97 T2w = T2t - T2v;
Chris@42 98 T35 = T2t + T2v;
Chris@42 99 T2k = T1l - T1r;
Chris@42 100 T1s = T1l + T1r;
Chris@42 101 T1v = T1t * T1u;
Chris@42 102 T2l = T1t * T1x;
Chris@42 103 }
Chris@42 104 T1A = cr[WS(rs, 11)];
Chris@42 105 T1D = ci[WS(rs, 11)];
Chris@42 106 T1z = W[20];
Chris@42 107 T1w = W[5];
Chris@42 108 T1C = W[21];
Chris@42 109 {
Chris@42 110 E T2m, T1y, T2o, T1E, T2n, T1B, T9;
Chris@42 111 Ta = cr[WS(rs, 4)];
Chris@42 112 T2n = T1z * T1D;
Chris@42 113 T1B = T1z * T1A;
Chris@42 114 T2m = FNMS(T1w, T1u, T2l);
Chris@42 115 T1y = FMA(T1w, T1x, T1v);
Chris@42 116 T2o = FNMS(T1C, T1A, T2n);
Chris@42 117 T1E = FMA(T1C, T1D, T1B);
Chris@42 118 Td = ci[WS(rs, 4)];
Chris@42 119 T9 = W[6];
Chris@42 120 T2p = T2m - T2o;
Chris@42 121 T36 = T2m + T2o;
Chris@42 122 T2r = T1E - T1y;
Chris@42 123 T1F = T1y + T1E;
Chris@42 124 Tb = T9 * Ta;
Chris@42 125 T1J = T9 * Td;
Chris@42 126 }
Chris@42 127 Tg = cr[WS(rs, 12)];
Chris@42 128 Tj = ci[WS(rs, 12)];
Chris@42 129 Tf = W[22];
Chris@42 130 Tc = W[7];
Chris@42 131 Ti = W[23];
Chris@42 132 }
Chris@42 133 }
Chris@42 134 {
Chris@42 135 E TQ, TT, TR, T2c, TW, TZ, TV, TS, TY;
Chris@42 136 {
Chris@42 137 E To, Tr, Tp, T1Q, Tu, Tx, Tt, Tq, Tw;
Chris@42 138 {
Chris@42 139 E T1K, Te, T1M, Tk, T1L, Th, Tn;
Chris@42 140 To = cr[WS(rs, 2)];
Chris@42 141 T1L = Tf * Tj;
Chris@42 142 Th = Tf * Tg;
Chris@42 143 T1K = FNMS(Tc, Ta, T1J);
Chris@42 144 Te = FMA(Tc, Td, Tb);
Chris@42 145 T1M = FNMS(Ti, Tg, T1L);
Chris@42 146 Tk = FMA(Ti, Tj, Th);
Chris@42 147 Tr = ci[WS(rs, 2)];
Chris@42 148 Tn = W[2];
Chris@42 149 T3k = T1K + T1M;
Chris@42 150 T1N = T1K - T1M;
Chris@42 151 T3z = Te - Tk;
Chris@42 152 Tl = Te + Tk;
Chris@42 153 Tp = Tn * To;
Chris@42 154 T1Q = Tn * Tr;
Chris@42 155 }
Chris@42 156 Tu = cr[WS(rs, 10)];
Chris@42 157 Tx = ci[WS(rs, 10)];
Chris@42 158 Tt = W[18];
Chris@42 159 Tq = W[3];
Chris@42 160 Tw = W[19];
Chris@42 161 {
Chris@42 162 E T1R, Ts, T1T, Ty, T1S, Tv, TP;
Chris@42 163 TQ = cr[WS(rs, 1)];
Chris@42 164 T1S = Tt * Tx;
Chris@42 165 Tv = Tt * Tu;
Chris@42 166 T1R = FNMS(Tq, To, T1Q);
Chris@42 167 Ts = FMA(Tq, Tr, Tp);
Chris@42 168 T1T = FNMS(Tw, Tu, T1S);
Chris@42 169 Ty = FMA(Tw, Tx, Tv);
Chris@42 170 TT = ci[WS(rs, 1)];
Chris@42 171 TP = W[0];
Chris@42 172 T1U = T1R - T1T;
Chris@42 173 T2W = T1R + T1T;
Chris@42 174 T1P = Ts - Ty;
Chris@42 175 Tz = Ts + Ty;
Chris@42 176 TR = TP * TQ;
Chris@42 177 T2c = TP * TT;
Chris@42 178 }
Chris@42 179 TW = cr[WS(rs, 9)];
Chris@42 180 TZ = ci[WS(rs, 9)];
Chris@42 181 TV = W[16];
Chris@42 182 TS = W[1];
Chris@42 183 TY = W[17];
Chris@42 184 }
Chris@42 185 {
Chris@42 186 E T13, T16, T14, T26, T19, T1c, T18, T15, T1b;
Chris@42 187 {
Chris@42 188 E T2d, TU, T2f, T10, T2e, TX, T12;
Chris@42 189 T13 = cr[WS(rs, 5)];
Chris@42 190 T2e = TV * TZ;
Chris@42 191 TX = TV * TW;
Chris@42 192 T2d = FNMS(TS, TQ, T2c);
Chris@42 193 TU = FMA(TS, TT, TR);
Chris@42 194 T2f = FNMS(TY, TW, T2e);
Chris@42 195 T10 = FMA(TY, TZ, TX);
Chris@42 196 T16 = ci[WS(rs, 5)];
Chris@42 197 T12 = W[8];
Chris@42 198 T2g = T2d - T2f;
Chris@42 199 T30 = T2d + T2f;
Chris@42 200 T25 = TU - T10;
Chris@42 201 T11 = TU + T10;
Chris@42 202 T14 = T12 * T13;
Chris@42 203 T26 = T12 * T16;
Chris@42 204 }
Chris@42 205 T19 = cr[WS(rs, 13)];
Chris@42 206 T1c = ci[WS(rs, 13)];
Chris@42 207 T18 = W[24];
Chris@42 208 T15 = W[9];
Chris@42 209 T1b = W[25];
Chris@42 210 {
Chris@42 211 E T27, T17, T29, T1d, T28, T1a, TA;
Chris@42 212 TB = cr[WS(rs, 14)];
Chris@42 213 T28 = T18 * T1c;
Chris@42 214 T1a = T18 * T19;
Chris@42 215 T27 = FNMS(T15, T13, T26);
Chris@42 216 T17 = FMA(T15, T16, T14);
Chris@42 217 T29 = FNMS(T1b, T19, T28);
Chris@42 218 T1d = FMA(T1b, T1c, T1a);
Chris@42 219 TE = ci[WS(rs, 14)];
Chris@42 220 TA = W[26];
Chris@42 221 T2a = T27 - T29;
Chris@42 222 T31 = T27 + T29;
Chris@42 223 T2h = T17 - T1d;
Chris@42 224 T1e = T17 + T1d;
Chris@42 225 TC = TA * TB;
Chris@42 226 T1X = TA * TE;
Chris@42 227 }
Chris@42 228 TH = cr[WS(rs, 6)];
Chris@42 229 TK = ci[WS(rs, 6)];
Chris@42 230 TG = W[10];
Chris@42 231 TD = W[27];
Chris@42 232 TJ = W[11];
Chris@42 233 }
Chris@42 234 }
Chris@42 235 }
Chris@42 236 {
Chris@42 237 E T2U, T3u, T2Z, T21, T1W, T34, T2X, T37, T3t, T3q, T3e, T32, T3i, T3h;
Chris@42 238 {
Chris@42 239 E T3f, T3r, T1H, T3s, TO, T3g;
Chris@42 240 {
Chris@42 241 E Tm, T1Y, TF, T20, TL, T3p, T1Z, TI;
Chris@42 242 T2U = T8 - Tl;
Chris@42 243 Tm = T8 + Tl;
Chris@42 244 T1Z = TG * TK;
Chris@42 245 TI = TG * TH;
Chris@42 246 T1Y = FNMS(TD, TB, T1X);
Chris@42 247 TF = FMA(TD, TE, TC);
Chris@42 248 T20 = FNMS(TJ, TH, T1Z);
Chris@42 249 TL = FMA(TJ, TK, TI);
Chris@42 250 T3p = T3k + T3o;
Chris@42 251 T3u = T3o - T3k;
Chris@42 252 {
Chris@42 253 E T1f, TM, T1G, T3j, T2V, TN;
Chris@42 254 T2Z = T11 - T1e;
Chris@42 255 T1f = T11 + T1e;
Chris@42 256 T21 = T1Y - T20;
Chris@42 257 T2V = T1Y + T20;
Chris@42 258 T1W = TF - TL;
Chris@42 259 TM = TF + TL;
Chris@42 260 T1G = T1s + T1F;
Chris@42 261 T34 = T1s - T1F;
Chris@42 262 T2X = T2V - T2W;
Chris@42 263 T3j = T2W + T2V;
Chris@42 264 T3f = T35 + T36;
Chris@42 265 T37 = T35 - T36;
Chris@42 266 T3t = Tz - TM;
Chris@42 267 TN = Tz + TM;
Chris@42 268 T3r = T1G - T1f;
Chris@42 269 T1H = T1f + T1G;
Chris@42 270 T3s = T3p - T3j;
Chris@42 271 T3q = T3j + T3p;
Chris@42 272 T3e = Tm - TN;
Chris@42 273 TO = Tm + TN;
Chris@42 274 T3g = T30 + T31;
Chris@42 275 T32 = T30 - T31;
Chris@42 276 }
Chris@42 277 }
Chris@42 278 cr[WS(rs, 12)] = T3r - T3s;
Chris@42 279 ci[WS(rs, 11)] = T3r + T3s;
Chris@42 280 ci[WS(rs, 7)] = TO - T1H;
Chris@42 281 T3i = T3g + T3f;
Chris@42 282 T3h = T3f - T3g;
Chris@42 283 cr[0] = TO + T1H;
Chris@42 284 }
Chris@42 285 {
Chris@42 286 E T3a, T2Y, T3x, T3v;
Chris@42 287 ci[WS(rs, 15)] = T3i + T3q;
Chris@42 288 cr[WS(rs, 8)] = T3i - T3q;
Chris@42 289 ci[WS(rs, 3)] = T3e + T3h;
Chris@42 290 cr[WS(rs, 4)] = T3e - T3h;
Chris@42 291 T3a = T2U + T2X;
Chris@42 292 T2Y = T2U - T2X;
Chris@42 293 T3x = T3u - T3t;
Chris@42 294 T3v = T3t + T3u;
Chris@42 295 {
Chris@42 296 E T2E, T1O, T3B, T3H, T2q, T2x, T3I, T23, T2R, T2O, T2J, T2K, T3C, T2H, T2B;
Chris@42 297 E T2j;
Chris@42 298 {
Chris@42 299 E T2F, T1V, T22, T2G;
Chris@42 300 {
Chris@42 301 E T3b, T33, T3c, T38;
Chris@42 302 T2E = T1I + T1N;
Chris@42 303 T1O = T1I - T1N;
Chris@42 304 T3b = T2Z - T32;
Chris@42 305 T33 = T2Z + T32;
Chris@42 306 T3c = T34 + T37;
Chris@42 307 T38 = T34 - T37;
Chris@42 308 T3B = T3z + T3A;
Chris@42 309 T3H = T3A - T3z;
Chris@42 310 {
Chris@42 311 E T3d, T3y, T3w, T39;
Chris@42 312 T3d = T3b + T3c;
Chris@42 313 T3y = T3c - T3b;
Chris@42 314 T3w = T38 - T33;
Chris@42 315 T39 = T33 + T38;
Chris@42 316 ci[WS(rs, 1)] = FMA(KP707106781, T3d, T3a);
Chris@42 317 cr[WS(rs, 6)] = FNMS(KP707106781, T3d, T3a);
Chris@42 318 ci[WS(rs, 13)] = FMA(KP707106781, T3y, T3x);
Chris@42 319 cr[WS(rs, 10)] = FMS(KP707106781, T3y, T3x);
Chris@42 320 ci[WS(rs, 9)] = FMA(KP707106781, T3w, T3v);
Chris@42 321 cr[WS(rs, 14)] = FMS(KP707106781, T3w, T3v);
Chris@42 322 cr[WS(rs, 2)] = FMA(KP707106781, T39, T2Y);
Chris@42 323 ci[WS(rs, 5)] = FNMS(KP707106781, T39, T2Y);
Chris@42 324 T2F = T1P + T1U;
Chris@42 325 T1V = T1P - T1U;
Chris@42 326 T22 = T1W + T21;
Chris@42 327 T2G = T1W - T21;
Chris@42 328 }
Chris@42 329 }
Chris@42 330 {
Chris@42 331 E T2M, T2N, T2b, T2i;
Chris@42 332 T2q = T2k - T2p;
Chris@42 333 T2M = T2k + T2p;
Chris@42 334 T2N = T2w + T2r;
Chris@42 335 T2x = T2r - T2w;
Chris@42 336 T3I = T22 - T1V;
Chris@42 337 T23 = T1V + T22;
Chris@42 338 T2R = FMA(KP414213562, T2M, T2N);
Chris@42 339 T2O = FNMS(KP414213562, T2N, T2M);
Chris@42 340 T2J = T25 + T2a;
Chris@42 341 T2b = T25 - T2a;
Chris@42 342 T2i = T2g + T2h;
Chris@42 343 T2K = T2g - T2h;
Chris@42 344 T3C = T2F - T2G;
Chris@42 345 T2H = T2F + T2G;
Chris@42 346 T2B = FMA(KP414213562, T2b, T2i);
Chris@42 347 T2j = FNMS(KP414213562, T2i, T2b);
Chris@42 348 }
Chris@42 349 }
Chris@42 350 {
Chris@42 351 E T2A, T3G, T2P, T2D, T3E, T3F, T3D, T2I;
Chris@42 352 {
Chris@42 353 E T24, T2L, T2C, T2y, T3J, T3L, T3K, T2S, T2z, T3M;
Chris@42 354 T2A = FNMS(KP707106781, T23, T1O);
Chris@42 355 T24 = FMA(KP707106781, T23, T1O);
Chris@42 356 T2S = FNMS(KP414213562, T2J, T2K);
Chris@42 357 T2L = FMA(KP414213562, T2K, T2J);
Chris@42 358 T2C = FMA(KP414213562, T2q, T2x);
Chris@42 359 T2y = FNMS(KP414213562, T2x, T2q);
Chris@42 360 T3J = FMA(KP707106781, T3I, T3H);
Chris@42 361 T3L = FNMS(KP707106781, T3I, T3H);
Chris@42 362 T2T = T2R - T2S;
Chris@42 363 T3K = T2S + T2R;
Chris@42 364 T3G = T2y - T2j;
Chris@42 365 T2z = T2j + T2y;
Chris@42 366 T3M = T2O - T2L;
Chris@42 367 T2P = T2L + T2O;
Chris@42 368 ci[WS(rs, 14)] = FMA(KP923879532, T3K, T3J);
Chris@42 369 cr[WS(rs, 9)] = FMS(KP923879532, T3K, T3J);
Chris@42 370 ci[0] = FMA(KP923879532, T2z, T24);
Chris@42 371 cr[WS(rs, 7)] = FNMS(KP923879532, T2z, T24);
Chris@42 372 cr[WS(rs, 13)] = FMS(KP923879532, T3M, T3L);
Chris@42 373 ci[WS(rs, 10)] = FMA(KP923879532, T3M, T3L);
Chris@42 374 T2D = T2B + T2C;
Chris@42 375 T3E = T2C - T2B;
Chris@42 376 }
Chris@42 377 T2Q = FNMS(KP707106781, T2H, T2E);
Chris@42 378 T2I = FMA(KP707106781, T2H, T2E);
Chris@42 379 T3F = FNMS(KP707106781, T3C, T3B);
Chris@42 380 T3D = FMA(KP707106781, T3C, T3B);
Chris@42 381 cr[WS(rs, 3)] = FMA(KP923879532, T2D, T2A);
Chris@42 382 ci[WS(rs, 4)] = FNMS(KP923879532, T2D, T2A);
Chris@42 383 cr[WS(rs, 1)] = FMA(KP923879532, T2P, T2I);
Chris@42 384 ci[WS(rs, 6)] = FNMS(KP923879532, T2P, T2I);
Chris@42 385 ci[WS(rs, 8)] = FMA(KP923879532, T3E, T3D);
Chris@42 386 cr[WS(rs, 15)] = FMS(KP923879532, T3E, T3D);
Chris@42 387 ci[WS(rs, 12)] = FMA(KP923879532, T3G, T3F);
Chris@42 388 cr[WS(rs, 11)] = FMS(KP923879532, T3G, T3F);
Chris@42 389 }
Chris@42 390 }
Chris@42 391 }
Chris@42 392 }
Chris@42 393 }
Chris@42 394 ci[WS(rs, 2)] = FMA(KP923879532, T2T, T2Q);
Chris@42 395 cr[WS(rs, 5)] = FNMS(KP923879532, T2T, T2Q);
Chris@42 396 }
Chris@42 397 }
Chris@42 398 }
Chris@42 399
Chris@42 400 static const tw_instr twinstr[] = {
Chris@42 401 {TW_FULL, 1, 16},
Chris@42 402 {TW_NEXT, 1, 0}
Chris@42 403 };
Chris@42 404
Chris@42 405 static const hc2hc_desc desc = { 16, "hf_16", twinstr, &GENUS, {104, 30, 70, 0} };
Chris@42 406
Chris@42 407 void X(codelet_hf_16) (planner *p) {
Chris@42 408 X(khc2hc_register) (p, hf_16, &desc);
Chris@42 409 }
Chris@42 410 #else /* HAVE_FMA */
Chris@42 411
Chris@42 412 /* Generated by: ../../../genfft/gen_hc2hc.native -compact -variables 4 -pipeline-latency 4 -n 16 -dit -name hf_16 -include hf.h */
Chris@42 413
Chris@42 414 /*
Chris@42 415 * This function contains 174 FP additions, 84 FP multiplications,
Chris@42 416 * (or, 136 additions, 46 multiplications, 38 fused multiply/add),
Chris@42 417 * 52 stack variables, 3 constants, and 64 memory accesses
Chris@42 418 */
Chris@42 419 #include "hf.h"
Chris@42 420
Chris@42 421 static void hf_16(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@42 422 {
Chris@42 423 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
Chris@42 424 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@42 425 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@42 426 {
Chris@42 427 INT m;
Chris@42 428 for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 30, MAKE_VOLATILE_STRIDE(32, rs)) {
Chris@42 429 E T7, T38, T1t, T2U, Ti, T37, T1w, T2R, Tu, T2t, T1C, T2c, TF, T2s, T1H;
Chris@42 430 E T2d, T1f, T1q, T2B, T2C, T2D, T2E, T1Z, T2k, T24, T2j, TS, T13, T2w, T2x;
Chris@42 431 E T2y, T2z, T1O, T2h, T1T, T2g;
Chris@42 432 {
Chris@42 433 E T1, T2T, T6, T2S;
Chris@42 434 T1 = cr[0];
Chris@42 435 T2T = ci[0];
Chris@42 436 {
Chris@42 437 E T3, T5, T2, T4;
Chris@42 438 T3 = cr[WS(rs, 8)];
Chris@42 439 T5 = ci[WS(rs, 8)];
Chris@42 440 T2 = W[14];
Chris@42 441 T4 = W[15];
Chris@42 442 T6 = FMA(T2, T3, T4 * T5);
Chris@42 443 T2S = FNMS(T4, T3, T2 * T5);
Chris@42 444 }
Chris@42 445 T7 = T1 + T6;
Chris@42 446 T38 = T2T - T2S;
Chris@42 447 T1t = T1 - T6;
Chris@42 448 T2U = T2S + T2T;
Chris@42 449 }
Chris@42 450 {
Chris@42 451 E Tc, T1u, Th, T1v;
Chris@42 452 {
Chris@42 453 E T9, Tb, T8, Ta;
Chris@42 454 T9 = cr[WS(rs, 4)];
Chris@42 455 Tb = ci[WS(rs, 4)];
Chris@42 456 T8 = W[6];
Chris@42 457 Ta = W[7];
Chris@42 458 Tc = FMA(T8, T9, Ta * Tb);
Chris@42 459 T1u = FNMS(Ta, T9, T8 * Tb);
Chris@42 460 }
Chris@42 461 {
Chris@42 462 E Te, Tg, Td, Tf;
Chris@42 463 Te = cr[WS(rs, 12)];
Chris@42 464 Tg = ci[WS(rs, 12)];
Chris@42 465 Td = W[22];
Chris@42 466 Tf = W[23];
Chris@42 467 Th = FMA(Td, Te, Tf * Tg);
Chris@42 468 T1v = FNMS(Tf, Te, Td * Tg);
Chris@42 469 }
Chris@42 470 Ti = Tc + Th;
Chris@42 471 T37 = Tc - Th;
Chris@42 472 T1w = T1u - T1v;
Chris@42 473 T2R = T1u + T1v;
Chris@42 474 }
Chris@42 475 {
Chris@42 476 E To, T1z, Tt, T1A, T1y, T1B;
Chris@42 477 {
Chris@42 478 E Tl, Tn, Tk, Tm;
Chris@42 479 Tl = cr[WS(rs, 2)];
Chris@42 480 Tn = ci[WS(rs, 2)];
Chris@42 481 Tk = W[2];
Chris@42 482 Tm = W[3];
Chris@42 483 To = FMA(Tk, Tl, Tm * Tn);
Chris@42 484 T1z = FNMS(Tm, Tl, Tk * Tn);
Chris@42 485 }
Chris@42 486 {
Chris@42 487 E Tq, Ts, Tp, Tr;
Chris@42 488 Tq = cr[WS(rs, 10)];
Chris@42 489 Ts = ci[WS(rs, 10)];
Chris@42 490 Tp = W[18];
Chris@42 491 Tr = W[19];
Chris@42 492 Tt = FMA(Tp, Tq, Tr * Ts);
Chris@42 493 T1A = FNMS(Tr, Tq, Tp * Ts);
Chris@42 494 }
Chris@42 495 Tu = To + Tt;
Chris@42 496 T2t = T1z + T1A;
Chris@42 497 T1y = To - Tt;
Chris@42 498 T1B = T1z - T1A;
Chris@42 499 T1C = T1y - T1B;
Chris@42 500 T2c = T1y + T1B;
Chris@42 501 }
Chris@42 502 {
Chris@42 503 E Tz, T1E, TE, T1F, T1D, T1G;
Chris@42 504 {
Chris@42 505 E Tw, Ty, Tv, Tx;
Chris@42 506 Tw = cr[WS(rs, 14)];
Chris@42 507 Ty = ci[WS(rs, 14)];
Chris@42 508 Tv = W[26];
Chris@42 509 Tx = W[27];
Chris@42 510 Tz = FMA(Tv, Tw, Tx * Ty);
Chris@42 511 T1E = FNMS(Tx, Tw, Tv * Ty);
Chris@42 512 }
Chris@42 513 {
Chris@42 514 E TB, TD, TA, TC;
Chris@42 515 TB = cr[WS(rs, 6)];
Chris@42 516 TD = ci[WS(rs, 6)];
Chris@42 517 TA = W[10];
Chris@42 518 TC = W[11];
Chris@42 519 TE = FMA(TA, TB, TC * TD);
Chris@42 520 T1F = FNMS(TC, TB, TA * TD);
Chris@42 521 }
Chris@42 522 TF = Tz + TE;
Chris@42 523 T2s = T1E + T1F;
Chris@42 524 T1D = Tz - TE;
Chris@42 525 T1G = T1E - T1F;
Chris@42 526 T1H = T1D + T1G;
Chris@42 527 T2d = T1D - T1G;
Chris@42 528 }
Chris@42 529 {
Chris@42 530 E T19, T1V, T1p, T22, T1e, T1W, T1k, T21;
Chris@42 531 {
Chris@42 532 E T16, T18, T15, T17;
Chris@42 533 T16 = cr[WS(rs, 15)];
Chris@42 534 T18 = ci[WS(rs, 15)];
Chris@42 535 T15 = W[28];
Chris@42 536 T17 = W[29];
Chris@42 537 T19 = FMA(T15, T16, T17 * T18);
Chris@42 538 T1V = FNMS(T17, T16, T15 * T18);
Chris@42 539 }
Chris@42 540 {
Chris@42 541 E T1m, T1o, T1l, T1n;
Chris@42 542 T1m = cr[WS(rs, 11)];
Chris@42 543 T1o = ci[WS(rs, 11)];
Chris@42 544 T1l = W[20];
Chris@42 545 T1n = W[21];
Chris@42 546 T1p = FMA(T1l, T1m, T1n * T1o);
Chris@42 547 T22 = FNMS(T1n, T1m, T1l * T1o);
Chris@42 548 }
Chris@42 549 {
Chris@42 550 E T1b, T1d, T1a, T1c;
Chris@42 551 T1b = cr[WS(rs, 7)];
Chris@42 552 T1d = ci[WS(rs, 7)];
Chris@42 553 T1a = W[12];
Chris@42 554 T1c = W[13];
Chris@42 555 T1e = FMA(T1a, T1b, T1c * T1d);
Chris@42 556 T1W = FNMS(T1c, T1b, T1a * T1d);
Chris@42 557 }
Chris@42 558 {
Chris@42 559 E T1h, T1j, T1g, T1i;
Chris@42 560 T1h = cr[WS(rs, 3)];
Chris@42 561 T1j = ci[WS(rs, 3)];
Chris@42 562 T1g = W[4];
Chris@42 563 T1i = W[5];
Chris@42 564 T1k = FMA(T1g, T1h, T1i * T1j);
Chris@42 565 T21 = FNMS(T1i, T1h, T1g * T1j);
Chris@42 566 }
Chris@42 567 T1f = T19 + T1e;
Chris@42 568 T1q = T1k + T1p;
Chris@42 569 T2B = T1f - T1q;
Chris@42 570 T2C = T1V + T1W;
Chris@42 571 T2D = T21 + T22;
Chris@42 572 T2E = T2C - T2D;
Chris@42 573 {
Chris@42 574 E T1X, T1Y, T20, T23;
Chris@42 575 T1X = T1V - T1W;
Chris@42 576 T1Y = T1k - T1p;
Chris@42 577 T1Z = T1X + T1Y;
Chris@42 578 T2k = T1X - T1Y;
Chris@42 579 T20 = T19 - T1e;
Chris@42 580 T23 = T21 - T22;
Chris@42 581 T24 = T20 - T23;
Chris@42 582 T2j = T20 + T23;
Chris@42 583 }
Chris@42 584 }
Chris@42 585 {
Chris@42 586 E TM, T1P, T12, T1M, TR, T1Q, TX, T1L;
Chris@42 587 {
Chris@42 588 E TJ, TL, TI, TK;
Chris@42 589 TJ = cr[WS(rs, 1)];
Chris@42 590 TL = ci[WS(rs, 1)];
Chris@42 591 TI = W[0];
Chris@42 592 TK = W[1];
Chris@42 593 TM = FMA(TI, TJ, TK * TL);
Chris@42 594 T1P = FNMS(TK, TJ, TI * TL);
Chris@42 595 }
Chris@42 596 {
Chris@42 597 E TZ, T11, TY, T10;
Chris@42 598 TZ = cr[WS(rs, 13)];
Chris@42 599 T11 = ci[WS(rs, 13)];
Chris@42 600 TY = W[24];
Chris@42 601 T10 = W[25];
Chris@42 602 T12 = FMA(TY, TZ, T10 * T11);
Chris@42 603 T1M = FNMS(T10, TZ, TY * T11);
Chris@42 604 }
Chris@42 605 {
Chris@42 606 E TO, TQ, TN, TP;
Chris@42 607 TO = cr[WS(rs, 9)];
Chris@42 608 TQ = ci[WS(rs, 9)];
Chris@42 609 TN = W[16];
Chris@42 610 TP = W[17];
Chris@42 611 TR = FMA(TN, TO, TP * TQ);
Chris@42 612 T1Q = FNMS(TP, TO, TN * TQ);
Chris@42 613 }
Chris@42 614 {
Chris@42 615 E TU, TW, TT, TV;
Chris@42 616 TU = cr[WS(rs, 5)];
Chris@42 617 TW = ci[WS(rs, 5)];
Chris@42 618 TT = W[8];
Chris@42 619 TV = W[9];
Chris@42 620 TX = FMA(TT, TU, TV * TW);
Chris@42 621 T1L = FNMS(TV, TU, TT * TW);
Chris@42 622 }
Chris@42 623 TS = TM + TR;
Chris@42 624 T13 = TX + T12;
Chris@42 625 T2w = TS - T13;
Chris@42 626 T2x = T1P + T1Q;
Chris@42 627 T2y = T1L + T1M;
Chris@42 628 T2z = T2x - T2y;
Chris@42 629 {
Chris@42 630 E T1K, T1N, T1R, T1S;
Chris@42 631 T1K = TM - TR;
Chris@42 632 T1N = T1L - T1M;
Chris@42 633 T1O = T1K - T1N;
Chris@42 634 T2h = T1K + T1N;
Chris@42 635 T1R = T1P - T1Q;
Chris@42 636 T1S = TX - T12;
Chris@42 637 T1T = T1R + T1S;
Chris@42 638 T2g = T1R - T1S;
Chris@42 639 }
Chris@42 640 }
Chris@42 641 {
Chris@42 642 E T1J, T27, T3a, T3c, T26, T3b, T2a, T35;
Chris@42 643 {
Chris@42 644 E T1x, T1I, T36, T39;
Chris@42 645 T1x = T1t - T1w;
Chris@42 646 T1I = KP707106781 * (T1C + T1H);
Chris@42 647 T1J = T1x + T1I;
Chris@42 648 T27 = T1x - T1I;
Chris@42 649 T36 = KP707106781 * (T2c - T2d);
Chris@42 650 T39 = T37 + T38;
Chris@42 651 T3a = T36 + T39;
Chris@42 652 T3c = T39 - T36;
Chris@42 653 }
Chris@42 654 {
Chris@42 655 E T1U, T25, T28, T29;
Chris@42 656 T1U = FNMS(KP382683432, T1T, KP923879532 * T1O);
Chris@42 657 T25 = FMA(KP382683432, T1Z, KP923879532 * T24);
Chris@42 658 T26 = T1U + T25;
Chris@42 659 T3b = T25 - T1U;
Chris@42 660 T28 = FMA(KP923879532, T1T, KP382683432 * T1O);
Chris@42 661 T29 = FNMS(KP923879532, T1Z, KP382683432 * T24);
Chris@42 662 T2a = T28 + T29;
Chris@42 663 T35 = T29 - T28;
Chris@42 664 }
Chris@42 665 cr[WS(rs, 7)] = T1J - T26;
Chris@42 666 cr[WS(rs, 11)] = T3b - T3c;
Chris@42 667 ci[WS(rs, 12)] = T3b + T3c;
Chris@42 668 ci[0] = T1J + T26;
Chris@42 669 ci[WS(rs, 4)] = T27 - T2a;
Chris@42 670 cr[WS(rs, 15)] = T35 - T3a;
Chris@42 671 ci[WS(rs, 8)] = T35 + T3a;
Chris@42 672 cr[WS(rs, 3)] = T27 + T2a;
Chris@42 673 }
Chris@42 674 {
Chris@42 675 E TH, T2L, T2W, T2Y, T1s, T2X, T2O, T2P;
Chris@42 676 {
Chris@42 677 E Tj, TG, T2Q, T2V;
Chris@42 678 Tj = T7 + Ti;
Chris@42 679 TG = Tu + TF;
Chris@42 680 TH = Tj + TG;
Chris@42 681 T2L = Tj - TG;
Chris@42 682 T2Q = T2t + T2s;
Chris@42 683 T2V = T2R + T2U;
Chris@42 684 T2W = T2Q + T2V;
Chris@42 685 T2Y = T2V - T2Q;
Chris@42 686 }
Chris@42 687 {
Chris@42 688 E T14, T1r, T2M, T2N;
Chris@42 689 T14 = TS + T13;
Chris@42 690 T1r = T1f + T1q;
Chris@42 691 T1s = T14 + T1r;
Chris@42 692 T2X = T1r - T14;
Chris@42 693 T2M = T2C + T2D;
Chris@42 694 T2N = T2x + T2y;
Chris@42 695 T2O = T2M - T2N;
Chris@42 696 T2P = T2N + T2M;
Chris@42 697 }
Chris@42 698 ci[WS(rs, 7)] = TH - T1s;
Chris@42 699 cr[WS(rs, 12)] = T2X - T2Y;
Chris@42 700 ci[WS(rs, 11)] = T2X + T2Y;
Chris@42 701 cr[0] = TH + T1s;
Chris@42 702 cr[WS(rs, 4)] = T2L - T2O;
Chris@42 703 cr[WS(rs, 8)] = T2P - T2W;
Chris@42 704 ci[WS(rs, 15)] = T2P + T2W;
Chris@42 705 ci[WS(rs, 3)] = T2L + T2O;
Chris@42 706 }
Chris@42 707 {
Chris@42 708 E T2f, T2n, T3g, T3i, T2m, T3h, T2q, T3d;
Chris@42 709 {
Chris@42 710 E T2b, T2e, T3e, T3f;
Chris@42 711 T2b = T1t + T1w;
Chris@42 712 T2e = KP707106781 * (T2c + T2d);
Chris@42 713 T2f = T2b + T2e;
Chris@42 714 T2n = T2b - T2e;
Chris@42 715 T3e = KP707106781 * (T1H - T1C);
Chris@42 716 T3f = T38 - T37;
Chris@42 717 T3g = T3e + T3f;
Chris@42 718 T3i = T3f - T3e;
Chris@42 719 }
Chris@42 720 {
Chris@42 721 E T2i, T2l, T2o, T2p;
Chris@42 722 T2i = FMA(KP382683432, T2g, KP923879532 * T2h);
Chris@42 723 T2l = FNMS(KP382683432, T2k, KP923879532 * T2j);
Chris@42 724 T2m = T2i + T2l;
Chris@42 725 T3h = T2l - T2i;
Chris@42 726 T2o = FNMS(KP923879532, T2g, KP382683432 * T2h);
Chris@42 727 T2p = FMA(KP923879532, T2k, KP382683432 * T2j);
Chris@42 728 T2q = T2o + T2p;
Chris@42 729 T3d = T2p - T2o;
Chris@42 730 }
Chris@42 731 ci[WS(rs, 6)] = T2f - T2m;
Chris@42 732 cr[WS(rs, 13)] = T3h - T3i;
Chris@42 733 ci[WS(rs, 10)] = T3h + T3i;
Chris@42 734 cr[WS(rs, 1)] = T2f + T2m;
Chris@42 735 cr[WS(rs, 5)] = T2n - T2q;
Chris@42 736 cr[WS(rs, 9)] = T3d - T3g;
Chris@42 737 ci[WS(rs, 14)] = T3d + T3g;
Chris@42 738 ci[WS(rs, 2)] = T2n + T2q;
Chris@42 739 }
Chris@42 740 {
Chris@42 741 E T2v, T2H, T32, T34, T2G, T2Z, T2K, T33;
Chris@42 742 {
Chris@42 743 E T2r, T2u, T30, T31;
Chris@42 744 T2r = T7 - Ti;
Chris@42 745 T2u = T2s - T2t;
Chris@42 746 T2v = T2r - T2u;
Chris@42 747 T2H = T2r + T2u;
Chris@42 748 T30 = Tu - TF;
Chris@42 749 T31 = T2U - T2R;
Chris@42 750 T32 = T30 + T31;
Chris@42 751 T34 = T31 - T30;
Chris@42 752 }
Chris@42 753 {
Chris@42 754 E T2A, T2F, T2I, T2J;
Chris@42 755 T2A = T2w + T2z;
Chris@42 756 T2F = T2B - T2E;
Chris@42 757 T2G = KP707106781 * (T2A + T2F);
Chris@42 758 T2Z = KP707106781 * (T2F - T2A);
Chris@42 759 T2I = T2w - T2z;
Chris@42 760 T2J = T2B + T2E;
Chris@42 761 T2K = KP707106781 * (T2I + T2J);
Chris@42 762 T33 = KP707106781 * (T2J - T2I);
Chris@42 763 }
Chris@42 764 ci[WS(rs, 5)] = T2v - T2G;
Chris@42 765 cr[WS(rs, 10)] = T33 - T34;
Chris@42 766 ci[WS(rs, 13)] = T33 + T34;
Chris@42 767 cr[WS(rs, 2)] = T2v + T2G;
Chris@42 768 cr[WS(rs, 6)] = T2H - T2K;
Chris@42 769 cr[WS(rs, 14)] = T2Z - T32;
Chris@42 770 ci[WS(rs, 9)] = T2Z + T32;
Chris@42 771 ci[WS(rs, 1)] = T2H + T2K;
Chris@42 772 }
Chris@42 773 }
Chris@42 774 }
Chris@42 775 }
Chris@42 776
Chris@42 777 static const tw_instr twinstr[] = {
Chris@42 778 {TW_FULL, 1, 16},
Chris@42 779 {TW_NEXT, 1, 0}
Chris@42 780 };
Chris@42 781
Chris@42 782 static const hc2hc_desc desc = { 16, "hf_16", twinstr, &GENUS, {136, 46, 38, 0} };
Chris@42 783
Chris@42 784 void X(codelet_hf_16) (planner *p) {
Chris@42 785 X(khc2hc_register) (p, hf_16, &desc);
Chris@42 786 }
Chris@42 787 #endif /* HAVE_FMA */